nengel@2: /*
nengel@2: * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
nengel@2: * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
nengel@2: *
nengel@2: * This file is part of FFmpeg.
nengel@2: *
nengel@2: * FFmpeg is free software; you can redistribute it and/or
nengel@2: * modify it under the terms of the GNU Lesser General Public
nengel@2: * License as published by the Free Software Foundation; either
nengel@2: * version 2.1 of the License, or (at your option) any later version.
nengel@2: *
nengel@2: * FFmpeg is distributed in the hope that it will be useful,
nengel@2: * but WITHOUT ANY WARRANTY; without even the implied warranty of
nengel@2: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
nengel@2: * Lesser General Public License for more details.
nengel@2: *
nengel@2: * You should have received a copy of the GNU Lesser General Public
nengel@2: * License along with FFmpeg; if not, write to the Free Software
nengel@2: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
nengel@2: */
nengel@2: #include "h264_types.h"
nengel@2: #include "h264_parser.h"
nengel@2: #include "h264_nal.h"
nengel@2: #include "h264_entropy.h"
nengel@2: #include "h264_rec.h"
nengel@2: #include "h264_pred_mode.h"
nengel@2: #include "h264_misc.h"
nengel@2: // #undef NDEBUG
nengel@2: #include <assert.h>
nengel@2: 
nengel@2: #pragma omp task inout(*pc, *nc) output(*sbe)
nengel@2: static void parse_task(H264Context *h, ParserContext *pc, NalContext *nc, SliceBufferEntry *sbe){
nengel@2:     H264Slice *s;
nengel@2: 
nengel@2:     if (!sbe->initialized){
nengel@2:         init_sb_entry(h, sbe);
nengel@2:         sbe->lines_total=h->mb_height;
nengel@2:     }
nengel@2: 
nengel@2:     av_read_frame_internal(pc, &sbe->gb);
nengel@2:     s = &sbe->slice;
nengel@2: 
nengel@2:     decode_nal_units(nc, s, &sbe->gb);
nengel@2: }
nengel@2: 
nengel@2: #pragma omp task inout(*ec) inout(*sbe)
nengel@2: static void decode_slice_entropy_task(H264Context *h, EntropyContext *ec, SliceBufferEntry *sbe){
nengel@2:     int i,j;
nengel@2:     H264Slice *s = &sbe->slice;
nengel@2:     GetBitContext *gb = &sbe->gb;
nengel@2:     H264Mb *mbs = sbe->mbs;
nengel@2: //     GetBitContext *gb = s->gb;
nengel@2:     CABACContext *c = &ec->c;
nengel@2: 
nengel@2:     if( !s->pps.cabac ){
nengel@2:         av_log(AV_LOG_ERROR, "Only cabac encoded streams are supported\n");
nengel@2:         return ;
nengel@2:     }
nengel@2: 
nengel@2:     init_dequant_tables(s, ec);
nengel@2:     ec->curr_qscale = s->qscale;
nengel@2:     ec->last_qscale_diff = 0;
nengel@2:     ec->chroma_qp[0] = get_chroma_qp((H264Slice *) s, 0, s->qscale);
nengel@2:     ec->chroma_qp[1] = get_chroma_qp((H264Slice *) s, 1, s->qscale);
nengel@2: 
nengel@2:     /* realign */
nengel@2:     align_get_bits( gb );
nengel@2:     /* init cabac */
nengel@2:     ff_init_cabac_decoder( c, gb->buffer + get_bits_count(gb)/8, (get_bits_left(gb) + 7)/8);
nengel@2: 
nengel@2:     ff_h264_init_cabac_states(ec, s, c);
nengel@2: 
nengel@2:     for(j=0; j<ec->mb_height; j++){
nengel@2:         init_entropy_buf(ec, s, j);
nengel@2:         for(i=0; i<ec->mb_width; i++){
nengel@2:             int eos,ret;
nengel@2:             H264Mb *m = &mbs[i + j*ec->mb_width];
nengel@2:             m->mb_x=i;
nengel@2:             m->mb_y=j;
nengel@2:             ec->m = m;
nengel@2: 
nengel@2:             ret = ff_h264_decode_mb_cabac(ec, s, c);
nengel@2:             eos = get_cabac_terminate( c);
nengel@2:             (void) eos;
nengel@2:             if( ret < 0 || c->bytestream > c->bytestream_end + 2) {
nengel@2:                 av_log(AV_LOG_ERROR, "error while decoding MB %d %d, bytestream (%td)\n", m->mb_x, m->mb_y, c->bytestream_end - c->bytestream);
nengel@2:                 return ;
nengel@2:             }
nengel@2:         }
nengel@2:     }
nengel@2: }
nengel@2: 
nengel@2: static void decode_super_mb_block(MBRecContext *d, H264Slice *s, SuperMBContext *smbc, H264Mb *mbs, int smb_x, int smb_y){
nengel@2:     MBRecState mrs;
nengel@2: //     memset(&mrs, 0, sizeof(MBRecState));
nengel@2: 
nengel@2:     for (int k=0, i= smb_y; i< smb_y + smbc->smb_height; i++, k++){
nengel@2:         init_mbrec_context(d, &mrs, s, i);
nengel@2:         for (int j= smb_x -k ; j< smb_x - k + smbc->smb_width; j++){
nengel@2:             if (i< d->mb_height && j >= 0 && j < d->mb_width){
nengel@2:                 h264_decode_mb_internal (d, &mrs, s, &mbs[i*d->mb_width+j]);
nengel@2:             }
nengel@2:         }
nengel@2:     }
nengel@2: }
nengel@2: 
nengel@2: #pragma omp task input(*d, *sbe, *ml, *mur) inout(*m)
nengel@2: static void decode_super_mb_task(MBRecContext *d, SliceBufferEntry *sbe, SuperMBContext *smbc, SuperMBTask *ml,
nengel@2: SuperMBTask *mur, SuperMBTask *m){
nengel@2:     H264Slice *s = &sbe->slice;
nengel@2:     H264Mb *mbs = sbe->mbs;
nengel@2:     decode_super_mb_block(d, s, smbc, mbs, m->smb_x, m->smb_y);
nengel@2: }
nengel@2: 
nengel@2: #pragma omp task input(*d, *sbe) inout(*sm)
nengel@2: static void draw_edges_task(MBRecContext *d, SliceBufferEntry *sbe, SuperMBContext *smbc, SuperMBTask *sm, int line){
nengel@2:     H264Slice *s = &sbe->slice;
nengel@2:     for (int i=line*smbc->smb_height; i< (line+1)*smbc->smb_height && i< d->mb_height; i++)
nengel@2:         draw_edges(d, s, i);
nengel@2: }
nengel@2: 
nengel@2: static void decode_mb_in_slice(H264Context *h, MBRecContext *d, SliceBufferEntry *sbe){
nengel@2:     int i,j;
nengel@2: 
nengel@2:     SuperMBContext *smbc = acquire_smbc(h);
nengel@2:     int smb_height =smbc->nsmb_height, smb_width= smbc->nsmb_width;
nengel@2:     SuperMBTask *smbs = smbc->smbs[0];
nengel@2: 
nengel@2:     SuperMBTask *sm=NULL, *sml, *smur;
nengel@2:     for(j=0; j< smb_height; j++){
nengel@2:         for(i=0; i< smb_width; i++){
nengel@2:             sm = smbs + j*smb_width + i;
nengel@2:             sml  = sm - ((i > 0) ? 1: 0);
nengel@2:             smur = sm + (((i < smb_width-1) && (j >0))  ? -smb_width+1: 0);
nengel@2:             decode_super_mb_task(d, sbe, smbc, sml, smur, sm);
nengel@2:         }
nengel@2:         draw_edges_task(d, sbe, smbc, sm, j);
nengel@2:     }
nengel@2:     #pragma omp taskwait on(*sm)
nengel@2: 
nengel@2:     release_smbc(h, smbc);
nengel@2: }
nengel@2: 
nengel@2: #pragma omp task inout(*d) inout(*sbe)
nengel@2: static void decode_slice_mb_task(H264Context *h, MBRecContext *d, SliceBufferEntry *sbe){
nengel@2:     H264Slice *s = &sbe->slice;
nengel@2: 
nengel@2:     for (int i=0; i<2; i++){
nengel@2:         for(int j=0; j< s->ref_count[i]; j++){
nengel@2:             if (s->ref_list_cpn[i][j] ==-1)
nengel@2:                 continue;
nengel@2:             int k;
nengel@2:             for (k=0; k< h->max_dpb_cnt; k++){
nengel@2:                 if(h->dpb[k].reference >= 2 && h->dpb[k].cpn == s->ref_list_cpn[i][j]){
nengel@2:                     s->dp_ref_list[i][j] = &h->dpb[k];
nengel@2:                     break;
nengel@2:                 }
nengel@2:             }
nengel@2:         }
nengel@2:     }
nengel@2: 
nengel@2:     #pragma omp critical (dpb)
nengel@2:     get_dpb_entry(h, s);
nengel@2: 
nengel@2:     if (!h->no_mbd){
nengel@2:         decode_mb_in_slice (h, d, sbe);
nengel@2:     }
nengel@2: 
nengel@2:     for (int i=0; i<s->release_cnt; i++){
nengel@2:         for(int j=0; j<h->max_dpb_cnt; j++){
nengel@2:             if(h->dpb[j].cpn== s->release_ref_cpn[i]){
nengel@2:                 #pragma omp critical (dpb)
nengel@2:                 release_dpb_entry(h, &h->dpb[j], 2);
nengel@2:                 break;
nengel@2:             }
nengel@2:         }
nengel@2:     }
nengel@2:     s->release_cnt=0;
nengel@2: }
nengel@2: 
nengel@2: // for static 3d wave
nengel@2: /*-------------------------------------------------------------------------------*/
nengel@2: #pragma omp task input(*d, *sbe, *ml, *mur, *mprev) inout(*m)
nengel@2: static void decode_3dwave_super_mb_task(MBRecContext *d, SliceBufferEntry *sbe, SuperMBContext *smbc, SuperMBTask *ml,
nengel@2: SuperMBTask *mur, SuperMBTask *mprev, SuperMBTask *m){
nengel@2:     H264Slice *s = &sbe->slice;
nengel@2:     H264Mb *mbs = sbe->mbs;
nengel@2: 
nengel@2:     decode_super_mb_block(d, s, smbc, mbs, m->smb_x, m->smb_y);
nengel@2: }
nengel@2: 
nengel@2: // int init_ref_count=0;
nengel@2: #pragma omp task inout(*d, *sbe, *init)
nengel@2: static void init_ref_list_and_get_dpb_task(H264Context *h, MBRecContext *d, SliceBufferEntry *sbe, int *init){
nengel@2:     H264Slice *s = &sbe->slice;
nengel@2:     for (int i=0; i<2; i++){
nengel@2:         for(int j=0; j< s->ref_count[i]; j++){
nengel@2:             if (s->ref_list_cpn[i][j] ==-1)
nengel@2:                 continue;
nengel@2:             int k;
nengel@2:             for (k=0; k<h->max_dpb_cnt; k++){
nengel@2:                 if(h->dpb[k].reference >= 2 && h->dpb[k].cpn == s->ref_list_cpn[i][j]){
nengel@2:                     s->dp_ref_list[i][j] = &h->dpb[k];
nengel@2:                     break;
nengel@2:                 }
nengel@2:             }
nengel@2:         }
nengel@2:     }
nengel@2: 
nengel@2:     #pragma omp critical (dpb)
nengel@2:     get_dpb_entry(h, s);
nengel@2: 
nengel@2: }
nengel@2: 
nengel@2: static SuperMBTask* add_decode_slice_3dwave_tasks(MBRecContext *d, SliceBufferEntry *sbe, SuperMBContext *smbc){
nengel@2:     int i,j;
nengel@2:     
nengel@2:     int smb_3d_height =smbc->nsmb_3dheight;
nengel@2:     int smb_height =smbc->nsmb_height, smb_width= smbc->nsmb_width;
nengel@2:     int smb_diff_prev = smb_height - smb_3d_height;
nengel@2:     SuperMBTask *sm=NULL, *sml, *smur, *smprev;
nengel@2: 
nengel@2:     SuperMBTask *smbs = smbc->smbs[smbc->index++]; smbc->index%=2; 
nengel@2:     SuperMBTask *smbs_prev = smbc->smbs[smbc->index]; // index rotates -> next == prev
nengel@2:     
nengel@2:     for(j=0; j<smb_3d_height ; j++){
nengel@2:         for(i=0; i< smb_width; i++){
nengel@2:             sm = smbs + j*smb_width + i;
nengel@2:             sml  = sm - ((i > 0) ? 1: 0);
nengel@2:             smur = sm + (((i < smb_width-1) && (j >0))  ? -smb_width+1: 0);
nengel@2:             smprev = smbs_prev + (j + smb_diff_prev+1)*smb_width -1;
nengel@2:             decode_3dwave_super_mb_task(d, sbe, smbc, sml, smur, smprev, sm);
nengel@2:         }
nengel@2:         draw_edges_task(d, sbe, smbc, sm, j);
nengel@2:     }
nengel@2: 
nengel@2:     for(; j< smb_height; j++){
nengel@2:         for(i=0; i< smb_width; i++){
nengel@2:             sm = smbs + j*smb_width + i;
nengel@2:             sml  = sm - ((i > 0) ? 1: 0);
nengel@2:             smur = sm + (((i < smb_width-1) && (j >0))  ? -smb_width+1: 0);
nengel@2:             decode_super_mb_task(d, sbe, smbc, sml, smur, sm);
nengel@2:         }
nengel@2:         draw_edges_task(d, sbe, smbc, sm, j);
nengel@2:     }
nengel@2:     return sm;
nengel@2: }
nengel@2: 
nengel@2: #pragma omp task inout(*d, *sbe, *release) input (*lastsmb)
nengel@2: static void release_ref_list_task(H264Context *h, SuperMBContext *smbc, MBRecContext *d, SliceBufferEntry *sbe, SuperMBTask *lastsmb, int *release){
nengel@2:     H264Slice *s = &sbe->slice;
nengel@2:     for (int i=0; i<s->release_cnt; i++){
nengel@2:         for(int j=0; j<h->max_dpb_cnt; j++){
nengel@2:             if(h->dpb[j].cpn== s->release_ref_cpn[i]){
nengel@2:                 #pragma omp critical (dpb)
nengel@2:                 release_dpb_entry(h, &h->dpb[j], 2);
nengel@2:                 break;
nengel@2:             }
nengel@2:         }
nengel@2:     }
nengel@2:     s->release_cnt=0;
nengel@2: 
nengel@2:     release_smbc(h, smbc);
nengel@2:     
nengel@2: }
nengel@2: 
nengel@2: // static void decode_mb_static_3dwave(H264Context *h, int mb_height, int mb_width, MBRecContext *d, H264Slice *s, H264Mb *mbs, SuperMBTask *smbs, SuperMBTask *smbs_prev){
nengel@2: //
nengel@2: // }
nengel@2: /*-------------------------------------------------------------------------------*/
nengel@2: //end for static 3d wave
nengel@2: 
nengel@2: #pragma omp task inout (*oc) input(*sbe)
nengel@2: static void output_task(H264Context *h, OutputContext *oc, SliceBufferEntry *sbe){
nengel@2:     DecodedPicture* out =output_frame(h, oc, sbe->slice.curr_pic, h->ofile, h->frame_width, h->frame_height);
nengel@2:     if (out){
nengel@2:         #pragma omp critical (dpb)
nengel@2:         release_dpb_entry(h, out, 1);
nengel@2:     }
nengel@2:     print_report(oc->frame_number, oc->video_size, 0, h->verbose);
nengel@2: }
nengel@2: 
nengel@2: /*
nengel@2: * The following code is the main loop of the file converter
nengel@2: */
nengel@2: //Put VMS entry point here
nengel@2: int h264_decode_ompss( H264Context *h) {
nengel@2:     const int bufs = h->pipe_bufs;
nengel@2: 
nengel@2:     ParserContext *pc;
nengel@2:     NalContext *nc;
nengel@2:     EntropyContext *ec[bufs];
nengel@2:     MBRecContext *rc[2];
nengel@2:     OutputContext *oc;
nengel@2:     SliceBufferEntry *sbe;
nengel@2:     SuperMBContext *smbc;
nengel@2: 
nengel@2:     DecodedPicture *out;
nengel@2:     int frames=0;
nengel@2: 
nengel@2: #if HAVE_LIBSDL2
nengel@2:     pthread_t sdl_thr;
nengel@2:     if (h->display){
nengel@2:         pthread_create(&sdl_thr, NULL, sdl_thread, h);
nengel@2:     }
nengel@2: #endif
nengel@2:     sbe= av_mallocz(sizeof(SliceBufferEntry) * bufs);
nengel@2: 
nengel@2: 
nengel@2:     pc = get_parse_context(h->ifile);
nengel@2:     nc = get_nal_context(h->width, h->height);
nengel@2: 
nengel@2:     for(int i=0; i<bufs; i++){
nengel@2:         ec[i] = get_entropy_context( h );
nengel@2:     }
nengel@2: 
nengel@2:     for(int i=0; i<2; i++){
nengel@2:         rc[i] = get_mbrec_context(h);
nengel@2:     }
nengel@2: 
nengel@2:     oc = get_output_context( h );
nengel@2: 
nengel@2:     av_start_timer();
nengel@2:     int k=0; int init, release;
nengel@2:     if (h->static_3d && bufs < h->num_frames ){
nengel@2:         int num_pre_ed =0;
nengel@2:         for (num_pre_ed=0; num_pre_ed< bufs -1 && !pc->final_frame; num_pre_ed++){
nengel@2:             parse_task( h, pc, nc, &sbe[k%bufs] );
nengel@2:             decode_slice_entropy_task(h, ec[k%bufs], &sbe[k%bufs]);
nengel@2:             #pragma omp taskwait on(*pc)
nengel@2:             k++;
nengel@2:         }
nengel@2: 
nengel@2:         while(!pc->final_frame && frames++ < h->num_frames && !h->quit){
nengel@2:             parse_task( h, pc, nc, &sbe[k%bufs] );
nengel@2:             decode_slice_entropy_task(h, ec[k%bufs], &sbe[k%bufs]);
nengel@2: 
nengel@2:             k++;
nengel@2: 
nengel@2:             init_ref_list_and_get_dpb_task(h, rc[k%2], &sbe[k%bufs], &init);
nengel@2:             smbc = acquire_smbc(h);
nengel@2:             SuperMBTask *lastsmb= add_decode_slice_3dwave_tasks(rc[k%2], &sbe[k%bufs], smbc);
nengel@2:             release_ref_list_task(h, smbc, rc[k%2], &sbe[k%bufs], lastsmb, &release);
nengel@2: 
nengel@2:             output_task (h, oc, &sbe[k%bufs]);
nengel@2:             #pragma omp taskwait on(*pc)
nengel@2:         }
nengel@2: 
nengel@2:         for (int i=0; i< num_pre_ed; i++){
nengel@2:             k++;
nengel@2:             init_ref_list_and_get_dpb_task(h, rc[k%2], &sbe[k%bufs], &init);
nengel@2:             smbc = acquire_smbc(h);
nengel@2:             SuperMBTask *lastsmb= add_decode_slice_3dwave_tasks(rc[k%2], &sbe[k%bufs], smbc);
nengel@2:             release_ref_list_task(h, smbc, rc[k%2], &sbe[k%bufs], lastsmb, &release);
nengel@2: 
nengel@2:             output_task (h, oc, &sbe[k%bufs]);
nengel@2:         }
nengel@2: 
nengel@2:     } else {
nengel@2:         while(!pc->final_frame && frames++ < h->num_frames && !h->quit){
nengel@2:             parse_task( h, pc, nc, &sbe[k%bufs] );
nengel@2: 
nengel@2:             decode_slice_entropy_task(h, ec[k%bufs], &sbe[k%bufs]);
nengel@2: 
nengel@2:             decode_slice_mb_task(h, rc[0], &sbe[k%bufs]);
nengel@2: 
nengel@2:             output_task (h, oc, &sbe[k%bufs]);
nengel@2:             #pragma omp taskwait on(*pc)
nengel@2:             k++;
nengel@2:         }
nengel@2:     }
nengel@2:     #pragma omp taskwait
nengel@2: 
nengel@2:     while ((out=output_frame(h, oc, NULL, h->ofile, h->frame_width, h->frame_height))) ;
nengel@2: 
nengel@2:     print_report(oc->frame_number, oc->video_size, 1, h->verbose);
nengel@2:     h->num_frames = oc->frame_number;
nengel@2:     /* finished ! */
nengel@2: 
nengel@2:     free_parse_context(pc);
nengel@2:     free_nal_context  (nc);
nengel@2:     free_output_context(oc);
nengel@2:     for (int i=0; i<bufs; i++){
nengel@2:         free_sb_entry(&sbe[i]);
nengel@2:         free_entropy_context(ec[i]);
nengel@2:     }
nengel@2:     av_free(sbe);
nengel@2: 
nengel@2:     for (int i=0; i<2; i++){
nengel@2:         free_mbrec_context(rc[i]);
nengel@2:     }
nengel@2: 
nengel@2: #if HAVE_LIBSDL2
nengel@2:     if (h->display){
nengel@2:         signal_sdl_exit(h);
nengel@2:         pthread_join(sdl_thr, NULL);
nengel@2:     }
nengel@2: #endif
nengel@2: 
nengel@2:     return 0;
nengel@2: }